@antoalphi I think you have already answered in the first line itself - Public Preview -- meaning not full developed for general/real use. Hence it comes with limitations or bugs which will be covered in General Available. Though still you may consider following points to tackle effectively:

• Group tables - split pipelines by schema / domain / size (not randomly)
• Use incremental (CDC) instead of full snapshot - reduces compute drastically
• Orchestrate pipelines (Databricks Workflows) - run sequentially or staggered to avoid many clusters at once
• Use serverless pipelines (where supported) - reduces cluster management overhead

Hope this helps, thanks.